Now that we have support for tensors with zero in its size, I believe it would be very handy to have support for accepting batches of size 0 in nn.functional functions.

A (non-exhaustive) list of functions that would be good supporting:

• conv{1-2-3}d
• conv_transpose{1-2-3}d
• batch_norm
• interpolate

Handling the losses is a bit trickier, because it generally involves computing a .mean(), which results in NaN due to 0 / 0 division. I'd expect having a 0 loss for empty batches to make sense, but that's debatable so might be worth postponing this decision.

cc @ezyang @gchanan @zou3519 @albanD @mruberry